Introduction
Moonshot AI announced Kimi K3 on July 16, 2026, and it immediately replaced K2.6 as the company's flagship. The Kimi API documentation now calls K3 "Kimi's most capable model to date" — the exact phrase it used for K2.6 three months ago — and the model is live across Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
The short version: 2.8 trillion parameters, a 1M-token context window, native vision, reasoning you cannot switch off, and an open-weights promise dated July 27, 2026 that has not landed yet.
A new architecture, not a bigger K2
K3 is not K2 scaled up. Moonshot built it on Kimi Delta Attention (KDA), described as a hybrid linear attention mechanism, together with Attention Residuals (AttnRes). Sparsity is handled by a Stable LatentMoE framework that activates just 16 of 896 experts per token — an unusually aggressive ratio for a Mixture-of-Experts model of this size.
Moonshot claims the combination yields "roughly 2.5x the overall scaling efficiency of K2." The full technical report is promised alongside the weights, so that claim is currently unverifiable. The announcement also names Quantile Balancing, Per-Head Muon, a Sigmoid Tanh Unit (SiTU), Gated MLA, and MXFP4 weights with MXFP8 activations.
For Kimi K2.6, the previous flagship, the comparison is stark: 1T parameters, 256K context, and optional thinking, against K3's 2.8T, 1M, and mandatory reasoning.
Thinking is no longer optional
This is the change most likely to bite existing integrations. K3 always has thinking mode enabled, and the top-level reasoning_effort field currently accepts only max — Moonshot says more levels are "coming soon."
Several sampling parameters are also fixed and must be omitted: temperature=1.0, top_p=0.95, n=1, and both penalties at zero. Prompt engineering is the only remaining lever.
There is a real cost to this. K2.7-Code's headline improvement was cutting "thinking-token usage by approximately 30%" — precisely the optimization K3 removes by reasoning at maximum effort on every call, including trivial ones.
Benchmarks, and an unusual concession
Moonshot published 35 benchmarks across six models — an unusually complete disclosure, and one that does not flatter K3. A selection, with K3 at reasoning_effort: max:
| Benchmark | Kimi K3 | Claude Fable 5 | GPT 5.6 Sol |
|---|---|---|---|
| BrowseComp | 91.2 | 88.0 | 90.4 |
| SWE Marathon | 42.0 | 35.0 | 39.0 |
| OmniDocBench | 91.1 | 89.8 | 85.8 |
| Terminal Bench 2.1 | 88.3 | 84.6 | 88.8 |
| GPQA-Diamond | 93.5 | 92.6 | 94.1 |
| MathVision | 94.3 | 94.8 | 95.8 |
| DeepSWE | 67.5 | 70.0 | 73.0 |
| HLE-Full | 43.5 | 53.3 | 44.5 |
Counted across the full table, K3 beats every listed competitor on 7 of 35 rows and loses to at least one on 27 — Claude Fable 5 takes 20 of those, GPT 5.6 Sol the rest. That is less damning than it sounds: against Claude Opus 4.8, GPT 5.5, and GLM-5.2 the model wins comfortably and often. The gap is specifically to the two frontier leaders, and K3's wins cluster in agentic search and document work rather than reasoning or vision.
The comparisons also are not like-for-like. Moonshot states that models are evaluated "under one of three agentic harnesses — KimiCode, Claude Code, or Codex," and many competitor scores are cited from third parties rather than measured: Program Bench numbers come from vals.ai, BrowseComp figures from Anthropic's and OpenAI's own posts, GDPval from Artificial Analysis. K3's Terminal Bench 88.3 is a KimiCode run while competitors get their best score across any harness.
The launch post is unusually candid about where K3 lands. It reports "frontier-level performance across our evaluation suite, consistently outperforming other tested models," then states plainly that K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol." A lab conceding second place in its own announcement is worth more than the table above it.
Moonshot lists three known limitations, and they are refreshingly concrete. Excessive proactiveness: on ambiguous intent, K3 "may make unexpected decisions on the user's behalf." Sensitivity to thinking history: K3 was trained with thinking history preserved, so if a harness drops it — or you switch a live session from another model to K3 — "generation quality may become highly unstable." And a plain admission that K3 "exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol."
The open-weights claim needs an asterisk
Moonshot calls K3 "the world's first open 3T-class model" and states that "the full model weights will be released by July 27, 2026."
As of July 17, 2026, no Kimi-K3 repository exists on Hugging Face, and Moonshot has not named a license. K2.6 and K2.7-Code ship under a Modified MIT License; nothing says K3 will follow.
So K3 today is a proprietary API model with an open-weights promise attached. The promise has a decent track record behind it — the K2 line is genuinely open, and Cursor's Composer 2.5 is post-trained on the open Kimi K2.5 checkpoint — but a date is a plan, not a fact. Nobody should architect around self-hosted K3 until the repository is real. And when it is, a 2.8T MoE will demand a far larger cluster than the 1T K2.6 that already needed a serious multi-GPU rig.
Pricing
Per 1M tokens, in USD:
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
kimi-k3 | $0.30 | $3.00 | $15.00 |
kimi-k2.6 | $0.16 | $0.95 | $4.00 |
kimi-k2.7-code | $0.19 | $0.95 | $4.00 |
Pricing is flat pay-as-you-go with no tiering by context length — the millionth token costs what the first does, which is a genuinely aggressive stance for a 1M-context large language model. Cache hits run one-tenth of a cache miss, so prompt caching is the first optimization worth making.
The rate card understates the real delta, though. K3 costs ~4x K2.6 per output token and burns reasoning tokens on every call. For sustained coding loops, kimi-k2.7-code at $4.00 output remains the cheaper tool.
The launch also started a clock on the older lineup: "Following the Kimi K3 launch, kimi-k2.5 and the moonshot-v1 series are no longer available to newly registered users (full platform sunset on August 31)." New accounts have already lost them; existing users have until that sunset, for which the docs state no year.
Conclusion
K3 is a real architectural step — new attention, a new MoE framework, a 4x context jump — and Moonshot is honest that it lands behind the leading proprietary models rather than ahead of them. The open-weights release on July 27 is the thing to watch: if a 2.8T model ships under a permissive license, it resets what "open-weight frontier model" means. Until then, K3 is an expensive API model that always thinks, and K2.6 remains the only Kimi flagship you can actually download.
Sources
- Kimi K3 announcement — Moonshot AI, July 16, 2026
- Kimi K3 quickstart guide — architecture,
reasoning_effort, fixed parameters - Kimi API model list — model IDs, context lengths, deprecations
- Kimi K3 pricing — per-token rates
- Kimi Code models — plan-gated K3 access
- Moonshot AI on Hugging Face — published weights, K3 absent as of July 17, 2026